Illustrative article · Sample editorial content and publication date, prepared for this website. Not a disclosure of proprietary research.
One layer, expanded inputs
A functional link network replaces hidden layers with a fixed expansion of its inputs. Each indicator is passed through a small set of nonlinear functions, often trigonometric terms, and a single layer of weights combines the expanded features. The network can represent curved relationships while keeping the number of trainable parameters small.
Solved, not searched
When the output weights are fitted in closed form, as in the extreme learning machine approach, training reduces to a regularised least-squares problem. There is no iterative descent to tune and no risk of stopping in a poor local minimum. Fitting takes a fraction of the time, which makes it practical to refit often and to test many variations out of sample.
Where it needs care
Efficiency is not the same as robustness. A wide expansion can fit noise as easily as structure, so the size of the expansion and the strength of regularisation matter more than the choice of network. Inputs such as moving averages, MACD, stochastic %K and %D, RSI and Williams %R are closely related to each other, and a signal is only as useful as its behaviour on data the model has never seen.